Optimization of sound fields reproduction based Higher-Order Ambisonics (HOA) using the Generative Adversarial Network (GAN)

Sound field reproduction using Higher-order Ambisonics (HOA) has many studies in recent years. However, in the HOA, sound fields are reproduced with the least square solution of spherical harmonics (SH) coefficients and not the global sound fields. In this paper, we try to reduce the reproduction error with a data-driven method. As we all known, the Generative Adversarial Networks (GAN) can be used to generate data similar to a data set. With the GAN, the target sound fields are converted to sound fields that can be reproduced accurately in the proposed approach. The data set of target sound fields is updated with the generated fields which have less reproduction error, and thus reproduction errors are reduced. We simulated the performance with four loudspeakers, sound fields of 4 orders SH coefficients are reproduced with GAN and HOA at 1000 Hz, with average reproduction errors of 0.3 and 0.6, respectively. Simulations show that the space between the least-square solution and the optimization solution is reduced with our method. Furthermore, the performances of HOA are optimized.